Towards Explainable Graph Embeddings for Gait Assessment Using Per-Cluster Dimensional Weighting.

Area of Science:

• Biomechanics
• Machine Learning
• Medical Informatics

Background:

• Gait analysis systems are improving, but their application in home-based healthcare is limited by the
• black box
• nature of current machine learning models, hindering healthcare professional trust and adoption.

Purpose of the Study:

• To develop an end-to-end pipeline for creating explainable gait pathology assessment systems.
• To address the lack of transparency in machine learning models used for gait analysis in clinical and home settings.

Main Methods:

• A bespoke Spatio-temporal Graph Convolutional Network (ST-GCN) and per-joint Principal Component Analysis (PCA) were used to generate graph feature embeddings.
• A novel semi-supervised weighting function was developed to quantify and rank important joint features for pathology description.
• K-means clustering was employed for gait pathology classification.

Main Results:

• The proposed method achieved classification accuracy improvements of 4.53% to 16% over the state of the art across three datasets.
• The system successfully identified and described 14 different simulated gait pathologies, from limping to ataxic gait.
• The generated latent graph embeddings enabled accurate and explainable descriptions of gait abnormalities.

Conclusions:

• The developed system offers a workable solution for at-home gait assessment applications.
• It provides accurate and explainable descriptions of gait abnormalities without requiring pre-labeled pathology data.
• This enhances the practical utility and trustworthiness of AI-driven gait analysis in healthcare.